Agents and the roster
Staff with a role, a set of tools and a line they never cross.
An agent in Delagents is not a chat window. It has a job description, the integrations it is allowed to touch, a budget, and rules about when to stop and ask. Hire from the roster or build your own.
Eight roles that earn their keep on day one.
Research analyst, support triager, bookkeeper, outreach writer, meeting prep, inbox triage, content repurposer, recruiting screener. Each ships with instructions, tool scopes and a public list of where it always stops.
- Every template is a public page: what it does, where it stops, what a day costs
- Add one to your staff with a click
- Edit the instructions and the policy; the floor stays
Build your own in six steps.
Identity, instructions, tools, guardrails, a test in a sandbox where side-effect tools are disabled, then publish. Versions are immutable once published and every delegation pins the version it ran on.
- Sandbox test on a sample brief before it joins the roster
- Versioned; roll back by re-pinning
- Track record per agent: delegations done, success rate, average cost
- Gmaildraft, labelMercer, Sol, Ines
- HubSpotread · write after sign-offMercer, Ada
- NotionreadAda, Sol
- StripereadInes, Nico
- Slackread #support · post after sign-offNico
Details
What is in the box.
Scopes per agent
A connection holds what the admin approved. Each agent gets a subset, and requests outside it fail closed.
Budgets per agent
A monthly cap per agent on top of the per-delegation cap. It stops at the limit and asks.
Allowed hours
Restrict an agent to business hours or a maintenance window. Runs outside pause until the window opens.
Pause and duplicate
Pause an agent to stop new delegations without cancelling running ones. Duplicate to fork a variant.
Usage you can read
Runs and spend over time, cache efficiency, tool error rate, per agent and per workspace.
Library, later
Publish templates to a shared library with the same disclosure. Community-contributed templates are on the roadmap.
Questions
Before you hand anything over.
Which models do agents run on?
Claude models through our gateway. The default is the most capable model for planning and drafting, with cheaper models for routing and classification. Business plans can bring their own keys.
Can an agent learn from my corrections?
Reviews and requests for changes are attached to the delegation and shape the next run's instructions. Customer data is not used to train models.
Can agents talk to each other?
Not in the preview. Hand-offs between agents are planned for teams at scale.
Hand it off.
Your first delegation takes a minute. Your first sign-off takes a second.